[flyingviper]

This source-specific helper split improves the complete conditional exponent saving from κ = 0.00310915222 to 0.00311024570, building on Ali Romman's #5713 at a5348f7143f95fe0046b92b96172e30763c7d731. The increase is approximately 0.0352%; this is a modest improvement and leaves the selected frontier unchanged.

The construction uses different legal combinations at the input and output boundaries. An existing rank-two meeting restores helper 708 and lets a new independent helper take over its continuation. An exact later relation reduces the output correction to three ordinary additions. The input fields are arbitrary, including those at zero-dimensional frames. There are no added paid recursive transports: per LOCAL, the complete signed target gain is H2 + H7 - H9 > 0, with all added ordinary work charged. The complete caller has 66,348 paid jobs and 12,976 productive fields. The raw crossing enclosure is (0.00311024572012, 0.00311024572013).

The focused package retains the exact predecessor export, source hashes and attribution, an explicit constructor, proof, pricing substitutions, independent evidence, and registered replay commands. No shared framework files, workflows or selected-result records change.

Validation

• Registered normal and optimized replays passed, comparing complete WORD, WITNESS and RESULT payloads: all 753 arbitrary-input columns, separate transpose, 2,917 native maps and absolute phases, owned lifetimes, full caller, two exact moment engines, compiler, and 54 strict plus four functional outer predicates.
• Ten malformed instruction/boundary controls and two invoice controls were rejected; inherited caller, phase, cutoff and BIT controls remain active.
• Separate agents checked the boundary proof and live-version identity, all scalar columns, source custody and pricing. The independent caller audit from Independently audit P104 callers with explicit owned lifetimes #5717 passed all 16 controls, with ordinary allowance 124,341,508 below 184,647,552.
• Committed-byte/source-history and official export-integrity checks passed at a1534ac45401c9a9c80c07aa65c89e00e33594f8. Both normal and optimized commands also passed from the final Git-free, read-only export with an empty PATH (268.656 seconds total); all 22 exported files, including the export manifest, remained byte-identical.

Reproduce with:

python3 -B scripts/research_package.py --package research/p104-two-boundary-helper-split
python3 -B scripts/check_research_commit.py --package research/p104-two-boundary-helper-split --check-source-history

Expected scientific marker: PASS_P104_TWO_BOUNDARY_SPLIT_KAPPA_00311024570.

Scope and known CI dependency

The inherited all-width native, owned erasure, compiler, precision/tape and analytic/recovery assumptions remain. Finite certificate checks and separate AI agents do not establish formal or independent human acceptance of the complete multiplication theorem.

At base df0137efd35e11b4445734bf612bdd4c7ebc330c, two unchanged documentation tests fail because the current-CI section omits its selected review and source-manifest links. The same two failures reproduce locally and in this PR's plan job: 449 tests, two failures, 21 skipped. The scientific replay job was skipped. Existing #5703 supplies that documentation fix. Local scientific success is not presented as hosted CI success.

Credit Ali Romman's temporal update, Joel Pulikkan/GamingPuzzled's #5684, the retained helper-split, caller, native and pricing authors, and Chris R/Clear1nsight's #5550 caller lineage with the maintainer's #5599 review. Prepared for flyingviper with substantial OpenAI Codex assistance, including independent verification and literature agents. Apache-2.0. Open next questions and concrete first steps are retained in the package.